| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
fix(cosyvoice): load audio with soundfile instead of torchaudio Co-authored-by: yeqm<yeqiangmao@huawei.com> # message auto-generated for no-merge-commit merge: !3081 merge fix/cosyvoice-soundfile-audio-load into master fix(cosyvoice): load audio with soundfile instead of torchaudio Created-by: ye_qm Commit-by: yeqm Merged-by: ascend-robot Description: ## What this PR does / why we need it? Fixes CosyVoice3 audio loading on Ascend environments. torchaudio==2.10.0 的 load() 会忽略 backend 参数，始终通过 TorchCodec 解码；而 PyPI 上的 TorchCodec wheel 是按 CUDA 版 torch 链接的（依赖 libnvrtc.so.13、libcudart.so.13、libc10_cuda.so、libtorch_cuda.so），昇腾环境的 CPU 版 torch 不存在这些库，导致预处理与训练数据加载均失败： OSError: libnvrtc.so.13: cannot open shared object file: No such file or directory 本 PR 将 3 处 torchaudio.load 调用替换为 soundfile 读取（soundfile 已在 examples/cosyvoice3/requirements.txt 中，无新增依赖）： - examples/cosyvoice3/preprocess/extract_embedding.py - examples/cosyvoice3/preprocess/extract_speech_token.py（原有的 backend='soundfile' 在 torchaudio 2.10 中已被忽略，实际不生效） - mindspeed_mm/fsdp/data/datasets/cosyvoice3/cosyvoice_dataset.py sf.read 返回 [time, channel]，通过转置保持原有 [channel, time] 布局，下游逻辑不变。 ## Does this PR introduce any user-facing change? Yes. CosyVoice3 用户现在可以在昇腾环境正常执行数据预处理与微调，无模型行为变化。 ## How was this patch tested? Tested on Ascend 910B2 with Python 3.12.14, CANN 9.1.0, torch==2.10.0 and torch-npu==2.10.0.post4: Related issue: https://gitcode.com/Ascend/MindSpeed-MM/issues/670 See merge request: Ascend/MindSpeed-MM!3081 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
glm5.3 flash preformance Co-authored-by: liuxi_<liuxi75@huawei.com> Co-authored-by: Super User<liuxi_> # message auto-generated for no-merge-commit merge: !3113 merge glm53_flash into master feat: glm5.3flash 性能优化 Created-by: liuxi_ Commit-by: liuxi_;Super User Merged-by: ascend-robot Description: ## What this PR does / why we need it? glm5.3flash 性能优化, A5适配 ## Does this PR introduce any user-facing change? 不涉及 ## How was this patch tested? 本地已验证精度/性能,长跑loss下降 相关用例在后续pr统一补充 See merge request: Ascend/MindSpeed-MM!3113 | 1 个月前 | |
docs:rename Ascend Extension for PyTorch Co-authored-by: z60112595<zhaotao68@h-partners.com> # message auto-generated for no-merge-commit merge: !2836 merge master into master docs:rename Ascend Extension for PyTorch Created-by: kali00 Commit-by: z60112595 Merged-by: ascend-robot Description: ## What this PR does / why we need it? https://gitcode.com/Ascend/MindSpeed-MM/issues/506 1、更新Ascend Extension for PyTorch名称为TorchNPU 2、补充快速入门环境最低配置 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2836 | 2 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
feat: support A5 performance test Co-authored-by: suis-moi<zhushixiang@huawei.com> # message auto-generated for no-merge-commit merge: !3111 merge master into master feat: support A5 performance test Created-by: suis-moi Commit-by: suis-moi Merged-by: ascend-robot Description: ## What this PR does / why we need it? add scripts to support A5 performance test ## Does this PR introduce any user-facing change? 不涉及 ## How was this patch tested? 本地已进行了性能测试 See merge request: Ascend/MindSpeed-MM!3111 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
添加LTX2.3 ReadMe文档 Co-authored-by: vasileone<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !3060 merge ltx2.3_push into master 添加LTX2.3 ReadMe文档 Created-by: vasileone Commit-by: vasileone Merged-by: ascend-robot Description: ## What this PR does / why we need it? Please describe the background and detailed changes of the PR. If it is a bugfix, please attach the related issue. ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!3060 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
docs:rename Ascend Extension for PyTorch Co-authored-by: z60112595<zhaotao68@h-partners.com> # message auto-generated for no-merge-commit merge: !2836 merge master into master docs:rename Ascend Extension for PyTorch Created-by: kali00 Commit-by: z60112595 Merged-by: ascend-robot Description: ## What this PR does / why we need it? https://gitcode.com/Ascend/MindSpeed-MM/issues/506 1、更新Ascend Extension for PyTorch名称为TorchNPU 2、补充快速入门环境最低配置 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2836 | 2 个月前 | |
feature: Add MiniMax-H3 FSDP2 finetuning for T2VA, FL2VA, and Ref2VA Co-authored-by: vasileone<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !3026 merge minimax_h3_adpter into master feature: Add MiniMax-H3 FSDP2 finetuning for T2VA, FL2VA, and Ref2VA Created-by: vasileone Commit-by: vasileone Merged-by: ascend-robot Description: ## What this PR does / why we need it? 本 PR 将 MiniMax-H3 的全参数 DiT 微调接入 MindSpeed-MM FSDP2 后端，支持 T2VA、FL2VA 和 Ref2VA 三类任务。 MiniMax-H3 的文本编码、Video VAE、Audio VAE 和数据预处理仍由 DiffSynth-Studio 离线完成；MindSpeed-MM 读取 DiffSynth-Studio 生成的预 计算 .pth 缓存并负责 DiT 训练，从而避免在训练路径中引入不必要的在线预处理依赖。 主要改动如下： - 新增 MiniMax-H3 T2VA、FL2VA、Ref2VA 的 FSDP2 全参数微调支持； - 新增 DiffSynth-Studio 预计算缓存的数据集插件； - 新增 Ref2VA 的 visual/audio reference anchor 处理； - 集成 NPU Flash Attention、RMSNorm、RoPE、SwiGLU 融合算子； - 支持 MindSpeed-MM 重计算和 Ulysses Context Parallel； - 提供 T2VA、FL2VA、Ref2VA 三套可直接启动的配置文件和脚本； - MiniMax-H3 README，补充环境安装、权重下载、缓存生成、微调和正式训练说明。 MiniMax-H3 DiT 结构及训练语义参考 DiffSynth-Studio： https://github.com/modelscope/DiffSynth-Studio/tree/0ad850953156c028c6334ea8c17c50032a1ef169/diffsynth ## Does this PR introduce any user-facing change? 是。 用户可通过以下脚本启动对应任务的 FSDP2 微调： - examples/minimax_h3/finetune_minimax_h3_t2va.sh - examples/minimax_h3/finetune_minimax_h3_fl2va.sh - examples/minimax_h3/finetune_minimax_h3_ref2va.sh 对应配置文件： - examples/minimax_h3/minimax_h3_t2va.yaml - examples/minimax_h3/minimax_h3_fl2va.yaml - examples/minimax_h3/minimax_h3_ref2va.yaml 用户文档路径： - examples/minimax_h3/README.md ## How was this patch tested? 以 DiffSynth-Studio 为参考实现，在相同预计算缓存、初始 Transformer 权重及固定随机种子条件下，进行 loss 精度对齐验证。 精度对齐使用 DiffSynth-Studio 基于 HQ-OpenHumanVid 生成的预计算 .pth 缓存。MindSpeed-MM 与 DiffSynth-Studio 训练时读取完全相同的缓存，并按固定路径顺序加载。 - FL2VA：使用 130 条真实视频样本，分辨率为 704 x 1280、243 帧；每条样本包含 prompt、目标视频及同步音频，并预计算首尾帧条件。 - Ref2VA：使用同一批 130 条真实视频样本。每条样本的 reference 由目标首帧图像、目标开头 1 秒的同步视频和音频构成。 ### 测试结果 | 任务 | 测试配置 | Loss 平均相对误差 | Loss 最大相对误差 | 结果 | | ------ | --------------------------------------------------------------------- | ----------------: | ----------------: | ------------------------ | | FL2VA | 5 个 DiT Block，500 step，8 NPU | 0.015047% | 0.090183% | 误差低于 1% | | Ref2VA | 5 个 DiT Block，500 step，8 NPU，Ulysses CP=2 | 0.039115% | 0.192966% | 误差低于 1% | ### GPU、NPU精度对齐实验  ### 性能表现 相同配置及数据下 - Diffsynth 源仓微调的单步耗时是33~34s - MindSpeed-MM 仓微调的单步耗时是29S (性能≈1.11 x GPU A100) 限制说明： - Ulysses CP 需要使用 flash_attention_2 或 flash_attention_3； - T2VA、FL2VA 必须使用 FL2VA transformer 权重；Ref2VA 必须使用 Ref2VA transformer 权重； - 不同任务类型、分辨率或帧数需要使用独立生成的 DiffSynth-Studio 缓存。 See merge request: Ascend/MindSpeed-MM!3026 | 1 个月前 | |
feat: optimize MiniMax-M3-VL FSDP2 training with NPU fused ops and Op Replay Co-authored-by: vasileone<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !3059 merge minimax_m3_push_new into master feat: optimize MiniMax-M3-VL FSDP2 training with NPU fused ops and Op Replay Created-by: vasileone Commit-by: vasileone;hanyyy Merged-by: ascend-robot Description: ## What this PR does / why we need it? 本 PR 对 MiniMax-M3-VL 的 FSDP2 训练实现进行了性能优化，并完善了面向开发者的训练配置和使用文档。 主要改动如下： - 接入融合版 clipped SwiGLU 算子，并保留 MiniMax-M3 的 swiglu_alpha 和 swiglu_limit 激活语义。 - 在专家并行调度过程中传递 MiniMax-M3 特有的 SwiGLU 参数。 - 优化稀疏注意力 Mask 的构造方式，使用基于视图的扩展替代重复的张量物化。 - 在视觉编码器中接入 NPU 融合旋转位置编码算子。 - 调整多模态特征提取与文本 Embedding 的执行顺序，并统一视觉特征和语言模型的计算数据类型。 - 在示例配置中显式启用 NPU 融合专家并行实现。 - 使用非重入式激活值重计算，并为融合注意力算子接入 Op Replay，减少反向阶段的重复计算。 - 调整 FSDP 包装范围，并将示例中的权重和数据路径修改为便于开源使用的相对路径。 - 完善 MiniMax-M3-VL 使用文档，包括环境准备、权重转换、数据准备、参数配置和分布式训练流程。 上述优化在保留现有 MiniMax-M3-VL 训练能力的基础上，降低了部分算子和重计算开销。 ## Does this PR introduce any user-facing change? 是。 MiniMax-M3-VL 示例配置包含以下变化： - 权重和数据集使用相对路径配置。 - 显式启用 NPU 融合专家并行实现。 - 使用非重入式激活值重计算。 - 默认为融合注意力算子启用 Op Replay。 - 支持配置激活值交换容量和预取策略。 开发者需要根据实际环境和训练任务调整权重路径、数据集路径、并行策略以及 Op Replay 的 capacity_mb。 本 PR 未移除或修改已有命令行接口。 相关文档和配置： - examples/minimax_m3_vl/README.md - examples/minimax_m3_vl/minimax_m3_config.yaml - examples/minimax_m3_vl/finetune_minimax_m3.sh ## How was this patch tested? 本 PR 通过以下方式进行验证： - 对 MiniMax-M3-VL 模型代码进行 Python 语法检查和模块导入检查。 - 验证 YAML 配置能够正常解析，并检查训练启动脚本语法。 - 验证融合 clipped SwiGLU 算子的计算结果以及 MiniMax-M3 特有的激活函数语义。 - 在相同权重、数据集、随机种子和并行配置下，对优化前后的 FSDP2 分布式训练进行对比。 - 在 8K 序列长度下完成超过 20 个共同训练 step 的 loss 对比，平均 loss 相对误差为 0.000872%，低于 1% 的精度误差要求。 - 排除预热 step 后，对稳态迭代耗时和设备峰值显存进行对比。 使用约束和注意事项： - Op Replay 的效果与序列长度、微批次大小、缓存算子以及 capacity_mb 密切相关，需要结合实际任务进行性能调优。 - cache_ops 并非配置得越多越好。只有当算子的重计算开销高于张量交换和内存管理开销时，缓存该算子才可能获得性能收益。 - 示例配置仅缓存融合注意力算子，可作为常用配置起点，但不保证适用于所有训练规格。 - 当前示例配置使用 Tensor Parallel Size 1，未启用 Context Parallel。 - 当前文档提供 Hugging Face 权重转换为 DCP 权重的流程。 See merge request: Ascend/MindSpeed-MM!3059 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
docs:rename Ascend Extension for PyTorch Co-authored-by: z60112595<zhaotao68@h-partners.com> # message auto-generated for no-merge-commit merge: !2836 merge master into master docs:rename Ascend Extension for PyTorch Created-by: kali00 Commit-by: z60112595 Merged-by: ascend-robot Description: ## What this PR does / why we need it? https://gitcode.com/Ascend/MindSpeed-MM/issues/506 1、更新Ascend Extension for PyTorch名称为TorchNPU 2、补充快速入门环境最低配置 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2836 | 2 个月前 | |
docs: 整改UserGuide/config中相关文档的说明 Co-authored-by: weixin_44031810<gaojie75@huawei.com> # message auto-generated for no-merge-commit merge: !3000 merge master into master docs: 整改UserGuide/config中相关文档的说明 Created-by: gaojie_ Commit-by: weixin_44031810 Merged-by: ascend-robot Description: ## What this PR does / why we need it? 本 PR 属于纯文档变更，目标修正 UserGuide/config 相关文档的说明内容。变更统计显示共修改 7 个文件、新增 1035 行、删除 0 行； 主要改动 修正 UserGuide/config 文档说明: 本次改动聚焦于更正 UserGuide/config 相关文档中的表述，属于文档说明层面的修正。 纯新增型文档变更: 变更以新增内容为主（新增 1035 行、删除 0 行），未涉及代码逻辑、接口、数据结构或配置项的实际修改。 https://gitcode.com/Ascend/MindSpeed-MM/issues/565 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!3000 | 1 个月前 | |
fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Co-authored-by: WendongPang<pangwendong@huawei.com> # message auto-generated for no-merge-commit merge: !3065 merge preload into master fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Created-by: WendongPang Commit-by: WendongPang Merged-by: ascend-robot Description: ## What this PR does / why we need it? 开启 enable_preload时，检查点保存时会出现 1 个 batch 的偏移。保存在检查点中的 DataLoader 状态是预取超前的位置，并非真实已消费的训练样本位置。因此恢复训练的第一步会使用第 k+1 个 batch，而非第 k 个 batch，造成损失、梯度范数与连续训练的结果不一致。 ## Does this PR introduce any user-facing change? 将以下几项默认置为false： training: no_load_optim: false # Do not load optimizer state; load by default. no_load_rng: false # Do not load RNG state; load by default. no_save_optim: false # Do not save optimizer state; save by default. no_save_rng: false # Do not save RNG state; save by default. ## How was this patch tested? 修正后开启preload断点续训精度对齐  See merge request: Ascend/MindSpeed-MM!3065 | 1 个月前 | |
fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Co-authored-by: WendongPang<pangwendong@huawei.com> # message auto-generated for no-merge-commit merge: !3065 merge preload into master fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Created-by: WendongPang Commit-by: WendongPang Merged-by: ascend-robot Description: ## What this PR does / why we need it? 开启 enable_preload时，检查点保存时会出现 1 个 batch 的偏移。保存在检查点中的 DataLoader 状态是预取超前的位置，并非真实已消费的训练样本位置。因此恢复训练的第一步会使用第 k+1 个 batch，而非第 k 个 batch，造成损失、梯度范数与连续训练的结果不一致。 ## Does this PR introduce any user-facing change? 将以下几项默认置为false： training: no_load_optim: false # Do not load optimizer state; load by default. no_load_rng: false # Do not load RNG state; load by default. no_save_optim: false # Do not save optimizer state; save by default. no_save_rng: false # Do not save RNG state; save by default. ## How was this patch tested? 修正后开启preload断点续训精度对齐  See merge request: Ascend/MindSpeed-MM!3065 | 1 个月前 | |
fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Co-authored-by: WendongPang<pangwendong@huawei.com> # message auto-generated for no-merge-commit merge: !3065 merge preload into master fix: Fix dataloader state mismatch with enable_preload for checkpoint resume Created-by: WendongPang Commit-by: WendongPang Merged-by: ascend-robot Description: ## What this PR does / why we need it? 开启 enable_preload时，检查点保存时会出现 1 个 batch 的偏移。保存在检查点中的 DataLoader 状态是预取超前的位置，并非真实已消费的训练样本位置。因此恢复训练的第一步会使用第 k+1 个 batch，而非第 k 个 batch，造成损失、梯度范数与连续训练的结果不一致。 ## Does this PR introduce any user-facing change? 将以下几项默认置为false： training: no_load_optim: false # Do not load optimizer state; load by default. no_load_rng: false # Do not load RNG state; load by default. no_save_optim: false # Do not save optimizer state; save by default. no_save_rng: false # Do not save RNG state; save by default. ## How was this patch tested? 修正后开启preload断点续训精度对齐  See merge request: Ascend/MindSpeed-MM!3065 | 1 个月前 | |
docs(qwen3_asr): add DCP to HF weight conversion guide Co-authored-by: ye_qm<yeqiangmao@huawei.com> # message auto-generated for no-merge-commit merge: !3107 merge docs/qwen3-asr-dcp2hf into master docs(qwen3_asr): add DCP to HF weight conversion guide Created-by: ye_qm Commit-by: ye_qm Merged-by: ascend-robot Description: ## What this PR does / why we need it? 背景：examples/qwen3_asr/README.md 目前仅覆盖 HF → DCP 的离线转换，缺少训练产物（DCP）转回 HuggingFace 格式的指引。用户在微调完成后无法从示例文档获得导出权重用于推理部署的路径。 变更内容：仅新增文档，无代码改动。在 examples/qwen3_asr/README.md 增加「3. 权重反向转换（DCP 转 HF）」章节： 1. 给出基于内置转换工具 mindspeed_mm.fsdp.tasks.checkpoint.converter dcp_to_hf 的完整命令，与 docs/zh/guides/development/checkpointer_guide.md 的官方流程保持一致； 2. 说明 dcp_dir 的两种目录布局（hf_to_dcp 转换产物的 release/ 与训练保存产物的 iter_xxxxxxx/）； 3. 说明 Qwen3-ASR 必须指定 --plugin mindspeed_mm/fsdp/models/qwen3_asr 的原因：qwen3_asr 这一 model_type 并非 transformers 内置架构，而是导入对应 mindspeed_mm 模型模块时才注册 AutoConfig/AutoProcessor；缺少该参数时转换会以 "Transformers does not recognize this architecture" 失败； 4. 标注转换产物中 config.json 为重新序列化生成（嵌套层 model_type 归一化、补充默认字段）等已知差异及影响范围。 ## Does this PR introduce any user-facing change? 不引入行为变更，仅补充文档。新增内容位于 examples/qwen3_asr/README.md「权重下载及转换 → 3. 权重反向转换（DCP 转 HF）」，并同步更新了该 README 的目录。文档告知了此前未记录的使用约束：转换 Qwen3-ASR 时 --plugin 参数为必选；纯 FSDP 环境需 NON_MEGATRON=true（与训练脚本一致）。 ## How was this patch tested? 在 Python 3.12、PyTorch 2.10.0、transformers 4.57.6 环境下完成本地验证： 1. 使用 Qwen3-ASR-1.7B 官方 config.json、processor/tokenizer 文件及包含 708 个权重键的两分片索引，构造确定性的轻量测试权重； 2. 按文档执行 hf_to_dcp，再使用新增命令执行 dcp_to_hf； 3. 转换成功，导出结果包含 708 个权重、2 个 safetensors 分片，逐 tensor 比较结果为 708/708 一致，0 mismatch； 4. 验证导出配置中的 model_type 分别为 qwen3_asr、qwen3_asr_thinker 和 qwen3_asr_text； 5. 负向验证确认 transformers 4.57.0 加载 Qwen3-ASR 插件时会因 check_model_inputs 接口不兼容而失败，升级至 4.57.6 后转换通过。 ## 关联 issue Closes #690 See merge request: Ascend/MindSpeed-MM!3107 | 1 个月前 | |
feat(qwen3omni): support A5 on qwen3omni Co-authored-by: yeqm<yeqiangmao@huawei.com> # message auto-generated for no-merge-commit merge: !3124 merge a5_qwen3omni into master feat(qwen3omni): support A5 on qwen3omni Created-by: ye_qm Commit-by: yeqm Merged-by: ascend-robot Description: ## What this PR does / why we need it? Add a single-node 8-NPU launch script for Qwen3-Omni training on A5. The script also collects the average iteration time and samples per second from the training log. ## Does this PR introduce any user-facing change? Yes. A new launch script is added: examples/qwen3omni/finetune_qwen3omni_v1_A5.sh Users can modify the distributed parameters and use it to launch Qwen3-Omni training on A5. ## How was this patch tested? Test on A5 See merge request: Ascend/MindSpeed-MM!3124 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat(torch): support fsdp_turbo low_precision Co-authored-by: yaoyaoxu<xuyaoyao.824404@huawei.com> # message auto-generated for no-merge-commit merge: !3095 merge support_fsdp_turbo_low_precision into master feat(torch): support fsdp_turbo low_precision Created-by: yaoyaoxu Commit-by: yaoyaoxu Merged-by: ascend-robot Description: ## What this PR does / why we need it? 由于mindspeed.fsdp低精模块不再演进，后续低精代码都在fsdp_turbo中更新，导致MM对低精的依赖需要从mindspeed.fsdp切换到fsdp_turbo。 本pr做了以下修改： 1.将量化及 optimizer hook 的依赖从 MindSpeed 内部实现切换至 fsdp_turbo 2.为重计算流程接入 FSDPTurbo 的 MXFP8 recompute context，保证 checkpoint 前向 / 反向阶段的量化状态正确 3.调整预取顺序、 纯fsdp2后端的Qwen3-VL MoE视觉 Flash Attention 的输入布局为TND，精度能对齐，且性能有1%+的收益 4.更新低精量化的文档 ## Does this PR introduce any user-facing change? 1.使用 FSDP2 低精度训练时，用户需要预先安装 FSDPTurbo 2.量化配置字段由旧格式迁移为新的 quantization_plan 配置格式，当前仅支持1*32量化 相关文档及示例： docs/zh/features/quantization.md examples/qwen3vl/qwen3vl_30B_config_v1_A5.yaml ## How was this patch tested? 1.recompute.py的修改已有ut看护：/tests/ut_fsdp/features/memory/test_act_stash.py 2.整网验证，开启重计算时，qwen3vl在all和on-demand两种低精通信方式下的loss是完全一致的 3.不安装 FSDPTurbo可以正常跑非低精的训练 See merge request: Ascend/MindSpeed-MM!3095 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3013 merge master into master docs: Update CANN version in legacy docs from 8.5.2 to 9.1.0 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 将遗留文档中的CANN版本信息从8.5.2更新到9.1.0 ## Does this PR introduce any user-facing change? / ## How was this patch tested? / See merge request: Ascend/MindSpeed-MM!3013 | 1 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
docs:rename Ascend Extension for PyTorch Co-authored-by: z60112595<zhaotao68@h-partners.com> # message auto-generated for no-merge-commit merge: !2836 merge master into master docs:rename Ascend Extension for PyTorch Created-by: kali00 Commit-by: z60112595 Merged-by: ascend-robot Description: ## What this PR does / why we need it? https://gitcode.com/Ascend/MindSpeed-MM/issues/506 1、更新Ascend Extension for PyTorch名称为TorchNPU 2、补充快速入门环境最低配置 ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2836 | 2 个月前 | |
feat: Bump PyTorch to 2.10.0 and Python to 3.12 Co-authored-by: zhangxubin<zhangxubin3@h-partners.com> # message auto-generated for no-merge-commit merge: !3020 merge master into master feat: Bump PyTorch to 2.10.0 and Python to 3.12 Created-by: MoCuishle-M Commit-by: zhangxubin Merged-by: ascend-robot Description: ## What this PR does / why we need it? 升级pytorch到2.10.0，python到3.12 mindspore相关readme没有更改 即将日落的verl相关模型readme没有更改 遗留： CI环境还未更新 到pytorch 2.10.0，所以当前pyproject.toml 采用灰度支持2.7.1的形式。 ## Does this PR introduce any user-facing change? 升级pytorch到2.10.0，python到3.12。 更新相关的readme ## How was this patch tested? 1. 本地切换CI环境验证通过 2. 手动验证关键模型qwen3.5、wan2.2 通过（详见PR https://gitcode.com/Ascend/MindSpeed-MM/pull/2976） 3. 部分模型测试用例（实际上为冒烟用例）更新到pytorch 2.10.0，测试通过 风险：没有验证到的场景与模型在pytorch 2.10.0、python 3.12可能存在问题 See merge request: Ascend/MindSpeed-MM!3020 | 1 个月前 | |
feat: support diffusers checkpointer Co-authored-by: chenpeizhe<chenpeizhe1@huawei.com> # message auto-generated for no-merge-commit merge: !3008 merge master into master feat: support diffusers checkpointer Created-by: chenpeizhe Commit-by: chenpeizhe Merged-by: ascend-robot Description: ## What this PR does / why we need it? Add diffusers checkpointer for generative models to transfer checkpoint format in FSDP2 backend. https://gitcode.com/Ascend/MindSpeed-MM/issues/576 ## Does this PR introduce any user-facing change? No, add diffusers checkpoiner. ## How was this patch tested? Test will be supported in others pr. See merge request: Ascend/MindSpeed-MM!3008 | 1 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 2 个月前 | ||
| 1 个月前 | ||
| 1 个月前 |